Context Rot
Context Rot refers to the phenomenon in which large language models or AI agents exhibit a gradual decline in attention to and utilization of earlier or critical information as context length continuously increases during task processing. Key characteristics include the model forgetting or neglecting important information in long conversations or documents, reduced reasoning accuracy as context accumulates, and the observation that condensed context often yields higher output quality. The concept reveals limitations of attention mechanisms in long-context processing.
Core Facts
Timeline (last 90 days)
LLMs exhibit Context Rot: the more tokens loaded, the lower the recall precision; the official recommendation is that a Skill exceeding 500 lines should be split, while practitioners tend to keep it within 200 lines.
Context Rot refers to the phenomenon where, after the context window is filled, the quality of an AI Agent's responses progressively deteriorates.
IBM points out that more context does not mean better performance, and that performance may degrade as context expands, a phenomenon referred to as Context Rot
All Facts (6)
Context Rot refers to the phenomenon where, as the number of conversation turns increases, irrelevant information accumulates and the attention budget becomes filled with low-value content, causing reasoning quality to continuously decline.
65%Unverified上下文腐烂（Context Rot）包含两种不同的故障模式：Distraction（注意力稀释）和Poisoning（因果链污染）
85%UnverifiedLLMs exhibit Context Rot: the more tokens loaded, the lower the recall precision; the official recommendation is that a Skill exceeding 500 lines should be split, while practitioners tend to keep it within 200 lines.
50%UnverifiedContext Rot refers to the phenomenon where, after the context window is filled, the quality of an AI Agent's responses progressively deteriorates.
50%UnverifiedIBM points out that more context does not mean better performance, and that performance may degrade as context expands, a phenomenon referred to as Context Rot
50%Unverified尽管上下文窗口很大，上下文腐化（Context Rot）依然是巨大问题，模型对上下文中不同位置信息的注意力分布不均匀（Lost in the Middle效应）
50%